Summary

Hermes routed provider=custom through chat_completions without sending temperature or parallel_tool_calls. Many local OpenAI-compatible stacks (e.g. llama.cpp / vLLM) then keep server defaults (often temperature=1.0) and only batch tool rounds when parallel_tool_calls is explicit.

Fixes #18470.

Changes

• ChatCompletionsTransport.build_kwargs
  When is_custom_provider is true:

  • If no fixed/overridden temperature applies, send temperature: 0.2 unless the user opts out (omit_temperature in compat options).
  • When tools are present, send parallel_tool_calls: true by default unless overridden.

• custom_providers (normalized in hermes_cli/config.py, surfaced via resolve_runtime_provider):

  • Optional: temperature (number), parallel_tool_calls (bool), omit_temperature: true.
  • Passed through CLI, gateway runtime dict, and AIAgent(custom_openai_request_options=...).

• Delegate: subagents inherit parent compat options when effective_provider == "custom".

Config example

custom_providers:
  - name: Local
    base_url: http://127.0.0.1:8080/v1
    model: your-model-alias
    temperature: 0.25
    parallel_tool_calls: true
    # omit_temperature: true   # uncomment to omit `temperature` and use server defaults